# US 10,872,225 B2 · Instantaneous search and comparison method for large-scale distributed palm vein micro-feature data

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 10,872,225 B2
- Document ID: US10872225B2
- Status: Granted
- Kind: B2
- Application number: 16/512,748
- Filed: 2019-07-16
- Published: 2020-10-29
- Granted: 2020-12-22
- Earliest claimed: 2019-04-23
- Assignees: Melux Technology Co., Ltd.
- Applicants: Melux Technology Co., Ltd.
- Inventors: Mengchun Yu; Qinglu Xie; Yongbo Xue
- Web page: https://patentgrep.com/patent/US10872225B2

## CPC

- G06V 10/431: {Frequency domain transformation; Autocorrelation}
- G06V 10/449: {Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters}
- G06V 30/19173: {Classification techniques}
- G06V 30/2504: {Coarse or fine approaches, e.g. resolution of ambiguities or multiscale approaches}
- G06V 40/1347: {Preprocessing; Feature extraction}
- G06V 40/1365: {Matching; Classification}
- G06V 40/50: Maintenance of biometric data or enrolment thereof
- G06V 40/14: Vascular patterns
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

The invention proposes an instantaneous search and comparison method for large-scale distributed palm vein micro-feature data, which consists of three parts: 1) feature extraction and calculation of palm vein micro-feature images; 2) building a feature database; 3) search and comparison. The technical solution provided by this invention, referring to the idea of GIS and web search method, is applied to the search and comparison of the palm vein micro-feature data, which enables instantaneous recognition on massive palm vein micro-feature data under large-scale, large-traffic and high-frequency application scenarios, and solves the technical problem that traditional palm vein recognition methods can not be applied to large-scale and large-traffic scenarios due to low speed thereof.

## Claims

### Claim 1

An instantaneous search and comparison method for large-scale distributed palm vein micro-feature data, includes the following steps:
S1: feature extraction and calculation of a palm vein micro-feature image, specifically comprises:
S1-1: building multi-scale palm vein feature spaces, let the original image be L0, building a downsample image pyramid sequence Sp={L0, L1, . . . , Ln-1}, where n is the total image number, c is the downsample factor;
S1-2: palm vein image enhancement, implement multi-scale Gauss filters on each layer of the sequence Sp, then extract the vein on the filtered images, combine the vein images under different scale kernels, obtaining the vein image sequence {V0, V1, . . . , Vn-1};
S1-3: extract the global and local features of the vein image sequence, combine the global and local features together to get the final feature F;
S2: building feature database, specifically comprises:
S2-1: bucket partition on the global features from the feature dataset, initialize z buckets randomly, classifying each global feature using cluster algorithm;
S2-2: building an inverted index for the local feature, building a “local feature ID—user ID” pair for the local feature of each feature F, building index system for all the local features, grouping the local features that are close in distance as one class, and putting the user IDs with the same local feature class to the inverted recording list;
S3: search and comparison, specifically comprises:
S3-1: introduce the feature H to be compared;
S3-2: get the bucket index sequence {id0, id1, . . . , idp-1} by using the p global features H0 of the feature H, then implement the search in the acquired p buckets parallelly, the searching process in each bucket comprises:
1) traverse all the local features of the feature H, find the t nearest neighbors of each local feature;
2) find the set Tm of user IDs corresponding to the t neighbors;
3) implement the intersection operation on all the IDs acquired from the m local features:
Gid=T0∩T1∩ . . . ∩Tm-1={ID0,ID1, . . . ,IDj}
obtaining the ID set of each bucket;
S3-3: implement the union operation on ID set of each bucket, i.e.,
G=G0∪G1∪ . . . Gp-1={g0,g1, . . . ,gj}
get the final ID set G;
S3-4: calculate the similarity score of feature H with each feature in set G;
s
=
∑
i
=
0
m
-
1
⁢

vg
i
-
v
⁢
p
i

m
where, vg is the local feature vector of G, vp is the local feature vector of H, m is the number of local features;
S3-5: select the one with the highest similarity score as the final matched target.

### Claim 2 (depends on 1)

The instantaneous search and comparison method for large-scale distributed palm vein micro-feature data according to claim 1, the multi-scale Gauss filters in step S1-2 are:
Zik(x,y)=G(x,y,kσ)*Li(x,y)
Li(x,y) denotes the input image, Zik(x,y) denotes the output image, * is the convolution operator, subscript i denotes the image sequence index, G(x,y,kσ) is the Gauss filter kernel, where:
G
⁡
(
x
,
y
,
k
⁢
⁢
σ
)
=
1
2
⁢
π
⁡
(
k
⁢
⁢
σ
)
2
⁢
exp
⁡
(
-
x
2
+
y
2
2
⁢
(
k
⁢
σ
)
2
)
k is the scale parameter of the Gauss filer kernel, and the parameter k∈K={1,2,3}.

### Claim 3 (depends on 1)

The instantaneous search and comparison method for large-scale distributed palm vein micro-feature data according to claim 1, extract vein on filtered image, and then combine the vein images under different scale kernels in the S1-2, specifically comprises:
first, performing vein splitting by using Phase Stretch Transform algorithm,
A[x,y]=∠IFFT2{{tilde over (K)}[α,β]·FFT2{B[x,y]}}
where, B[x,y] is the input image, A[x,y] is the output image, ∠<⋅> is the angle operator, FFT2 is the two-dimensional fast fourier transform, IFFT2 is the inverse two-dimensional fast fourier transform, α and β are the two-dimensional frequency variables, {tilde over (K)}[α,β] is a phase distortion kernel function, i.e.,
K
~
⁡
[
α
,
β
]
=
e
j
·
φ
⁡
[
α
,
β
]
φ
⁡
[
α
,
β
]
=
⁢
φ
polar
⁡
[
r
,
θ
]
=
⁢
φ
polar
⁡
[
r
]
=
⁢
S
·
W
·
r
·
tan
-
1
⁡
(
W
·
r
)
-
(
1
/
2
)
·
ln
⁡
(
1
+
(
W
·
r
)
2
)
W
·
r
max
·
tan
-
1
⁡
(
W
·
r
max
)
-
(
1
/
2
)
·
ln
⁡
(
1
+
(
W
·
r
max
)
2
)
where, r=√{square root over (α2+β2)}, the final value of φ is irrelevant to θ, thus can be omitted directly, rmax is the maximum value of r, S is the strength coefficient, and W is the distortion strength, the values of S and W depend on the image;
second, merging the vein images under different scale kernels,
V
i
=
merge
k
⁡
(
V
i
k
)
where, the merging operation consists of the following steps:
a) binarize Vik;
b) calculate the summation of the binarized Vik in corresponding position, i.e.,
V
i
⁡
(
x
,
y
)
=
∑
k
⁢
V
i
k
⁡
(
x
,
y
)
c) perform the region extremum growth on the image Vi, i.e.,
v
⁡
(
x
,
y
)
=
{
0
if
⁢
⁢
v
⁡
(
x
,
y
)
=
0
v
⁡
(
x
,
y
)
if
⁢
⁢
v
⁡
(
x
,
y
)
≠
0
⁢
⁢
and
⁢
⁢
v
max
⁢
-
⁢
neighbor
≤
thr
v
max
⁢
-
⁢
neighbor
if
⁢
⁢
v
⁡
(
x
,
y
)
≠
0
⁢
⁢
and
⁢
⁢
v
max
⁢
-
⁢
neighbor
>
thr
Where, vmax-neighbor is the maximum value in the region, thr is an experience threshold, and generating the vein image sequence (V0, V1, . . . , Vn-1) after merging.

### Claim 4 (depends on 1)

The instantaneous search and comparison method for large-scale distributed palm vein micro-feature data according to claim 1, the global and local features extraction for vein image sequence in S1-3, specifically comprises:
select the first p sequence images O={V0, V1, . . . , Vp-1} from the vein image sequence {V0, V1, . . . , Vn-1} for the global feature calculation, the remaining images used for the local feature calculation;
the network used for the global feature calculation is GlobalNet; primary vein extraction on the sequence images O one by one, resize the primary vein sequence images to the same size and input them to the network GlobalNet, generating the feature vector sequence:
FO={f0,f1, . . . ,fp-1};
region segmentation on the vein image and segmentation on the remaining images are required for the local feature calculation, suppose the number of remaining images is q, the whole palm is divided into m local regions, for each local region, input the local vein image into the network LocalNet and generate the local feature vector vmq, where the superscript q denotes the sequence index, subscript m denotes the local feature index in the whole palm, subsequently, merging the local features in the sequence with the same position, i.e.,
v
m
=
∑
i
=
0
q
-
1
⁢
v
m
i
q
get the final local feature Fl={v0, v1, . . . , vm-1};
combine the global and local features together to obtain the final feature F, i.e.,
F=(p,m,Fo,Fl).

### Claim 5 (depends on 1)

The instantaneous search and comparison method for large-scale distributed palm vein micro-feature data according to claim 1, classify each global feature using a cluster algorithm in S2-1, specifically comprises:
for the feature F, since there are p groups of global features, generating partition vector D={id0, id1, . . . , idp-1}, where id denotes the bucket index of the class k+1, counting the class index assigned to the feature F, the feature F will be considered to be the class with the most assignments, if all the assignments are 1, select the first group with index 0 as the feature F class id.